matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-11 00:07:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

update llama.cpp submodule to 77d1ac7 (#3030)

Browse Source
This commit is contained in:
Jeffrey Morgan
2024-03-09 15:55:34 -08:00
committed by GitHub
parent 0a7844413c
commit 1ffb1e2874
6 changed files with 24 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
llm/patches/02-cudaleaks.diff
View File

@@ -1,10 +1,10 @@
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index f255ad76..5b83acb1 100644
index b14cca61..02bfd4b1 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -28,6 +28,10 @@
#include <thread>
@@ -29,6 +29,10 @@
#include <signal.h>
#include <memory>
+#ifdef GGML_USE_CUBLAS
+extern "C" GGML_CALL void ggml_free_cublas(void);
@@ -13,7 +13,7 @@ index f255ad76..5b83acb1 100644
using json = nlohmann::json;
bool server_verbose = false;
@@ -648,6 +652,10 @@ struct server_context {
@@ -664,6 +668,10 @@ struct server_context {
llama_free_model(model);
model = nullptr;
}
@@ -24,7 +24,7 @@ index f255ad76..5b83acb1 100644
}
bool load_model(const gpt_params & params_) {
@@ -3339,6 +3347,7 @@ int main(int argc, char ** argv) {
@@ -3499,6 +3507,7 @@ int main(int argc, char ** argv) {
sigemptyset (&sigint_action.sa_mask);
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
@@ -33,10 +33,10 @@ index f255ad76..5b83acb1 100644
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 72bcec8c..50a45e3d 100644
index c207ff87..945708a4 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -43,6 +43,7 @@
@@ -46,6 +46,7 @@
#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
#define cublasCreate hipblasCreate
@@ -44,7 +44,7 @@ index 72bcec8c..50a45e3d 100644
#define cublasGemmEx hipblasGemmEx
#define cublasGemmBatchedEx hipblasGemmBatchedEx
#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
@@ -8751,10 +8752,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
@@ -8014,10 +8015,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
return g_cublas_loaded;
}
@@ -58,7 +58,7 @@ index 72bcec8c..50a45e3d 100644
#ifdef __HIP_PLATFORM_AMD__
// Workaround for a rocBLAS bug when using multiple graphics cards:
@@ -8764,7 +8765,7 @@ GGML_CALL void ggml_init_cublas() {
@@ -8027,7 +8028,7 @@ GGML_CALL void ggml_init_cublas() {
#endif
if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
@@ -67,7 +67,7 @@ index 72bcec8c..50a45e3d 100644
g_cublas_loaded = false;
fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
return;
@@ -8835,7 +8836,7 @@ GGML_CALL void ggml_init_cublas() {
@@ -8098,7 +8099,7 @@ GGML_CALL void ggml_init_cublas() {
// configure logging to stdout
// CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
@@ -76,11 +76,12 @@ index 72bcec8c..50a45e3d 100644
g_cublas_loaded = true;
}
}
@@ -12490,3 +12491,22 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
@@ -11753,3 +11754,23 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
}
return device_count;
}
+
+
+extern "C" GGML_CALL void ggml_free_cublas(void);
+GGML_CALL void ggml_free_cublas(void) {
+ for (int id = 0; id < g_device_count; ++id) {